Local Knowledge Base (RAG)

1. What is RAG and Why ZeroChat is Unique?

RAG (Retrieval-Augmented Generation) is a method that allows language models to ground their answers in your personal documents, manuals, codebases, or technical notes.

While typical SaaS tools require uploading sensitive files to remote servers, ZeroChat executes all document ingestion, chunking, and search indexing locally inside your browser using the Orama search engine backed by IndexedDB.

2. Knowledge Branches

To keep distinct projects cleanly separated, ZeroChat organizes knowledge into independent Branches:

• Create topic-specific branches (e.g., "API Specs", "Financial Reports", "Research Papers").
• Activate multiple branches from the RAG button in the composer toolbar. To create branches, upload, or manage documents, open RAG from the general settings menu.
• When combining branches in different languages, ZeroChat shows a notice without changing active branches.
• When a branch is active, the chat engine and agents gain direct retrieval access to its documents.

3. Document Ingestion & Chunking

Open RAG from the general settings menu, then drop files into its management panel or select them from disk:

• PDFs & Books: Parses text per page and captures embedded figures and diagrams.
• Markdown & Plain Text: Respects heading structures (#, ##) for semantic boundary division.
• Source Code Files: Groups code blocks respecting function and class scopes so technical context is preserved.
• Intelligent Chunking: Divides content into optimal token windows with configured overlap to ensure complete contextual coherence.
• Duplicate Detection: When uploading a file with the same name and path as an existing document in the branch, ZeroChat prompts you to either Replace the previous document or Ignore the duplicate. You can check the box to apply this decision to all remaining duplicates in the batch.
• Cooperative Cancellation: During long or multi-file ingestions, you can click the Stop button to halt the process cleanly at any point without corrupting the index or previously saved documents.

4. Retrieval Strategies & Agent Tools

The model queries active documents through RAG tools:

1. Active Agent Tool Calling: The model dynamically queries the database using specialized tools:
   • list_documents: Discovers available documents in the branch.
   • search_knowledge_base: Searches for specific terms and semantic matches.
   • read_knowledge_chunk: Reads a specific section in high resolution.
   • read_knowledge_image: Inspects embedded figures or chart images visually.

5. Configuration & Recommendations

IndexedDB + Orama

Files, metadata, and chunks are stored in IndexedDB. Orama builds an in-memory search index for the active branch.

No permissions or ingestion costs

Selecting a file only grants read access to that specific upload. No folder permissions are requested, and no LLM calls are made to summarize or index documents.

Configuring the model for document queries

For factual queries over documents, use a low temperature (0 to 0.2): this reduces variation and increases the likelihood that the model sticks to retrieved passages. Higher temperatures (0.7+) are better suited for creative writing tasks, not for extracting numbers or facts. The model must also support tool calling; if it doesn't, it won't be able to query this knowledge base.

Agent checkpoint for low-context models

Compact models or those with limited context windows can become overwhelmed when accumulating multiple document fragments. Enable the Agent Checkpoint (agent_checkpoint) from the Reasoning menu: it allows the model to consolidate intermediate findings and maintain a clear research plan. Especially recommended for models with fewer than 14B parameters or context windows under 32k tokens.